Mem Cogn
DOI 10.3758/s13421-014-0396-z
Task influences on the production and comprehension
of compound words
Niels Janssen & Petra E. Pajtas & Alfonso Caramazza
# Psychonomic Society, Inc. 2014
Abstract Studies of compound word processing have revealed effects of the compound’s constituents in a wide variety of word production and comprehension tasks. Surprisingly, effects of the compound’s constituents were not found in a
recent word production study using the picture-naming task.
Here, we examined whether these contrasting constituent
effects reflect methodological differences or whether they
reflect the influence of task-specific characteristics on compound processing. A large set of compounds was used to elicit
picture-naming latencies in Experiment 1 and visual lexical
decision latencies in Experiment 2. Regression analyses revealed surface frequency effects in picture naming and lexical
decision and constituent effects only in lexical decision. These
results rule out methodological differences as the basis for the
contrasting constituent effects observed between the various
tasks and imply that constituent effects are best understood as
arising out of task-specific characteristics.
Keywords Language production . Language
comprehension . Morphology . Compounds
The study of familiar compound words (e.g., doghouse) has
been used in the fields of word comprehension and production
to gain insight into the role of morphology in the mental
lexicon. The general finding is that effects of the compound’s
constituents have been observed in a variety of word
N. Janssen (*)
Cognitive Neuroscience and Psycholinguistics Lab,
Universidad de La Laguna, La Laguna 38205, Spain
e-mail: njanssen@ull.es
P. E. Pajtas : A. Caramazza
Harvard University, 33 Kirkland Street, Cambridge,
MA 02138, USA
A. Caramazza
Center for Mind/Brain Sciences, University of Trento, Trento, Italy
comprehension tasks, such as lexical decision and word reading (e.g., Andrews, 1986; De Jong, Feldman, Schreuder,
Pastizzo, & Baayen, 2002; Duñabeitia, Perea, & Carreiras,
2007; Juhasz, Starr, Inhoff, & Placke, 2003; Kuperman,
Schreuder, Bertram, & Baayen, 2009; Libben, Gibson, Yoon,
& Sandra, 2003; Pollatsek, Hyönä, & Bertram, 2000), and in
word production tasks, such as response association and picture–word interference (e.g., Bien, Levelt, & Baayen, 2005;
Dohmes, Zwitserlood, & Bölte, 2004; Gumnior, Bölte, &
Zwitserlood, 2006; Koester & Schiller, 2008; Roelofs, 1996;
Zwitserlood, Bölte, & Dohmes, 2000). In light of these data, a
rather surprising result was recently reported by Janssen, Bi,
& Caramazza (2008). That study found that compound picture
naming did not yield the constituent effects found in other
tasks. Here, we examined whether these contrasting data
reflect methodological differences between the studies or
whether they reflect task-specific influences on the processing
of compounds. To address this issue, we investigated compound
word processing in picture-naming and lexical decision tasks.
The majority of studies on compound word processing have
been conducted in the domain of visual word recognition. To
examine the role of the compound’s constituents in lexical
processing, these studies rely on a manipulation of measures
associated with the compound’s whole word and constituents.
Many studies have revealed effects of the lexical frequency of
the compound’s constituents and its surface (whole-word)
form. For example, Juhasz et al. (2003) asked English participants to perform naming, lexical decision, and reading in
sentence context tasks on visually presented compound words
while their eye movements and reaction times (RTs) were
registered. The frequency of the first and second constituents
of the compounds was varied factorially, while the compound’s surface frequency was held constant. The results revealed effects of the second constituent frequency on both eye
fixations and latencies in all three tasks. Similarly, Duñabeita
et al. (2007) found that lexical decision times were affected by
Mem Cogn
the frequency of the second constituent in both headfinal Spanish and head-initial Basque compound words.
Finally, Pollatsek et al. (2000) revealed that both surface
frequency and second constituent frequency affected eye
fixations in a reading in sentence context task with Finnish
compound words.1
The impact of morphology on lexical processing has also
been examined using other measures than lexical frequency.
For example, Kuperman et al. (2009) conducted a large scale
study involving 2,500 compounds in which Dutch participants
performed a lexical decision task while their eye movements
and decision latencies were registered. Besides examining the
impact of the compound’s surface and constituent frequency
on the dependent measures, they also considered variables
related to the compound’s constituent family size.
The variable constituent family size of a compound is
defined as the cardinality of the set of compounds that share
a constituent with a given compound in a given position.
Thus, for example, for the following set of compounds {windmill, windshield, windbag, windhose}, the left constituent
family size would be 4. In a regression analysis that relied
on a mixed effect technique, Kuperman et al. (2009) reported
that lexical decision latencies were sensitive to the compound’s surface frequency, to constituent frequency, and to
the compounds’ constituent family size measures (see also De
Jong et al., 2002; Juhasz & Berkowitz, 2011).
Although different models of complex word recognition
have been proposed to account for these results (e.g.,
Caramazza Laudanna, & Romani, 1988; Seidenberg &
Gonnerman, 2000; Taft, 2003; see Diependaele, Grainger, &
Sandra, 2012, for a recent review), a shared assumption in all
these models is that in the course of compound word recognition, the compound’s constituents are activated and influence word comprehension.
Studies from the field of word production have also found
evidence for a role of the compound’s constituents in processing (Bien et al., 2005; Roelofs, 1996). For example, Bien et al.
(2005) found constituent effects in compound word production using the response association task. In this task, participants first learn to associate a number of cues (i.e., other
words, spatial positions on the computer screen) with visually
or auditorily presented compound words. In the actual experiment, the cues are then used to elicit overt production of the
associated compound word. Bien et al. found that production
latencies were sensitive to various measures (i.e., frequency,
family size) associated with the compound’s constituents.
Other word production studies using the picture–word
interference task have yielded comparable results (Dohmes
1
There is disagreement in the word recognition literature about the
presence of first-constituent effects (cf. Juhasz et al., 2003; Taft &
Forster, 1976). Here, we simply wish to highlight that effects of the
constituents are found.
et al., 2004; Gumnior et al., 2006; Koester & Schiller, 2008;
Zwitserlood, Bölte & Dohmes, 2000, 2002). For example,
Zwitserlood et al. (2000) reported shorter naming latencies
to pictures with monomorphemic names (e.g., dog) when they
were preceded by a visually presented compound word prime
that contained the picture name (e.g., doghouse). On the basis
of these data, the authors argued in favor of a model that
assumes that a compound’s constituent morphemes play a role
in word production.
The overall conclusion that emerges from this brief review
of word comprehension and production studies is that compound processing yields constituent effects. Given this evidence, it is surprising that rather different results were recently
obtained by Janssen et al. (2008) using a picture-naming task.
Janssen et al. (2008) asked Mandarin Chinese and English
speakers to name pictures with compound names. The compound’s surface frequency and constituent frequencies were
orthogonally manipulated in a factorial design. The results
revealed that in both Chinese and English, naming latencies
were determined by the compound’s surface frequency, but
not by its constituent frequencies. Four additional control
experiments ruled out the possibility that these results were
artifacts related to picture recognition or articulatory processes. The absence of constituent effects in compound picture
naming are surprising, considering the success of the task in
elucidating how other components, such as semantics and
phonology, play a role in the production of words (e.g., Glaser,
1992). In other words, the data reported by Janssen et al.
(2008) raise the question of why the task would be specifically
insensitive to constituent effects in compound word
production.
Thus, whereas effects of the compound’s constituent morphemes have been repeatedly reported in many studies using a
variety of tasks, they were not found in the picture-naming
study by Janssen et al. (2008). This contrast suggests that taskspecific characteristics may have an important influence on
how compound words are processed. However, before such a
conclusion can be accepted, two issues need to be addressed.
First, the empirical evidence for the absence of constituent
effects in the picture-naming task is rather weak; it relies on a
single study. Consequently, establishing the reliability of the
picture-naming data reported by Janssen et al. (2008) is crucial. Second, there are many methodological differences between the previously reported studies and that of Janssen et al.
(2008). The studies used different languages, materials, and
analyses, and all of these might impact the results reported for
compound word processing. In order to satisfactorily address
the difference in findings, it is important that we examine
critically whether the contrasting results can be reduced to
methodological differences.
The two experiments reported below were designed to
address these issues. In Experiment 1, we attempted to establish the reliability of the picture-naming results. The
Mem Cogn
experimental design differed from the factorial design used by
Janssen et al. (2008). Native English speakers named a large
set of pictures (N = 150) with compound names whose surface
and constituent measures varied (see below for details). Each
participant named each picture twice—once before and once
after a picture familiarization stage. This was done to address
the possibility that the results observed by Janssen et al.
(2008) were due to the familiarization procedure. In Experiment 2, we examined visual lexical decision latencies from the
English Lexicon Project (Balota et al., 2007) for a subset of
the compounds used in Experiment 1. We focused on the
lexical decision task since this task has a long history in
revealing robust constituent effects. In both experiments, we
relied on the same analytical techniques to determine the
presence of constituent effects.
Thus, in Experiments 1 and 2, comparable sets of compounds from the same language were analyzed in the same
way. If the contrasting constituent effects observed in the
literature arose as a consequence of differences in languages,
stimulus sets, or analyses, we would not expect contrasting
constituent effects between Experiments 1 and 2.
Experiment 1: Picture naming
Method
Participants
Twenty-eight native English speakers who were students at
Harvard University participated in the experiment. All participants received course credit or compensation ($10 per hour).
Materials and design
All stimuli were selected from the database described by
Janssen, Pajtas, and Caramazza (2011). One hundred and fifty
pictures with compound names were used. They were black
and white line drawings that fit on a white square of 300 × 300
pixels. All compounds had two major constituents.
The 150 pictures were assembled into two blocks of 150
trials each. The manipulation of block served to examine the
effect of familiarization, where naming in block 1 occurred
before familiarization and naming in block 2 occurred after
familiarization. Each picture appeared once per block. The order
of trials was pseudorandomized such that, on successive trials,
picture names were semantically and phonologically unrelated.
In the regression analyses reported below, we distinguished
between critical and control variables. The critical variables
were related to the compounds’ surface and constituent measures. For each compound, we computed surface and constituent (left and right) measures using CELEX data (Baayen,
Piepenbrock, & van Rijn, 1993). Specifically, the compound’s
surface frequency was the lemma frequency (stem plus inflectional variants) of the whole compound word. Its left lemma
frequency was the lemma frequency of the left constituent, and
its left cumulative frequency was the left constituent’s lemma
frequency plus the sum of all the frequencies of complex words
in which the constituent appeared, taking into account homophonic variants of the constituent (i.e., homophone frequency).
Other measures were based on the compound’s constituent
family (De Jong et al., 2002; Schreuder & Baayen, 1997),
which is the set of compounds (or other derived words) that
share a constituent (e.g., {windmill, wind chill, windbag,
windshield, windy}). The left constituent family size was the
cardinality of the compound’s left constituent family; its left
positional frequency was the sum of the lemma frequencies of
the family members, and its left positional entropy was a
measure of the distribution of the relative frequencies of the
family members using Shannon’s entropy (Shannon, 2001).
The compound’s left complement frequency referred to the
sum of the frequencies of all complex, noncompound words in
which the constituent appeared, and Shannon’s entropy of this
set was called its left derivational entropy. Right constituent
measures were computed accordingly. Table 1 presents a
summary of the distribution of these variables. Finally, the
variable block denoted the number of times a picture was
named (i.e., before and after familiarization).
In addition, we included a number of control variables that
are known to affect latencies in psycholinguistic experiments in
general, as well as the picture-naming task in particular. First,
we considered the impact of the variable trial (the ordinal
position of an item in the experiment). In addition, we examined the impact of subjective familiarity, name–image agreement, and visual complexity ratings that were obtained for these
pictures (see Janssen et al., 2011, for details). The latter variables are typically associated with picture recognition processes
(e.g., Snodgrass & Vanderwart, 1980). We also considered the
variable compound word length in phonemes (Bertram &
Hyönä, 2003; Juhasz, 2008). Finally, the impact of three articulatory factors (i.e., voicing, plosiveness, fricativeness) was
taken into account, given the observation that such factors affect
latencies in word production tasks (e.g., Balota, Cortese,
Sergent-Marshall, Spieler, & Yap, 2004).
Procedure
The experiment was presented using DMDX (Forster &
Forster, 2003). In the first block, participants named a picture
and, subsequently, saw its name printed on the screen. Participants were told to use this name for the particular picture.
This was considered the familiarization stage. On each trial, a
fixation cross was presented (700 ms), followed by a blank
screen (200 ms), the presentation of the picture (2,000 ms), the
picture’s name (1,000 ms), and a blank screen (1,000 ms). In
the second block participants were instructed to name the
Mem Cogn
Table 1 Distribution of the variables associated with the compound’s left and right constituents in Experiments 1 and 2
Left Constituent
Right Constituent
Variables
Median
Range
Median
Range
Experiment 1
Lemma frequency
Cumulative frequency
Family size
Positional frequency
Positional entropy
Complement frequency
Derivational entropy
894
1,675
7
68
309,396
29
258,307
0–14,935
0–111,823
1–63
0–3,038
0–579,678
0–5984
0–555,677
621
1,575
13
77
603,204
34
278,545
4–29,231
0–82,033
1–426
0–6,600
0–2,698,981
0–7,182
0–642,252
Experiment 2
Lemma frequency
Cumulative frequency
Family size
Positional frequency
Positional entropy
Complement frequency
Derivational entropy
1,102
1,840
10
130
198,000
49
280,804
4–12,983
14–111,823
1–63
1–3,038
0–500,174
0–4,510
0–555,677
586
1,716
16
114
310,730
45
271,775
4–29,231
6–82,033
1–426
0–6,600
0–2,698,981
0–7,182
0–597,187
pictures again, and were told the picture’s name would not be
presented, and to use the name presented in the previous part.
In this block, each trial began with a fixation cross (700 ms),
followed by a blank screen (200 ms), the picture (2,000 ms),
and another blank screen (1,000 ms).
Analysis and results
Trials on which participants produced incorrect responses, as
well as trials on which the naming latency was less than
300 ms or greater than 2,500 ms, were discarded (1,940 trials
out of 8,400, or 23 %). The resulting set of 6,460 data points
was analyzed using linear mixed-effects methodology
(Baayen, 2008). Continuous predictors, except for RT, were
log-transformed to reduce skewness.
Statistical considerations
The full model that included all the critical and control variables described above suffered from high multicollinearity
(condition index = 48). High collinearity in regression models
is problematic for two reasons. First, high collinearity undermines a clear interpretation of the potential effects of predictors in the model, and second, high collinearity can lead to
problems in the stability of the model-fitting procedures. We
addressed the issue of collinearity with two measures.
First, all continuous variables were centered on their sample
means. Centering of variables both improves the interpretability
of variables in the model and reduces collinearity when predictors participate in interactions with other predictors. Second,
correlated variables were residualized. Residualizing in regression designs has the same effect as controlling for a variable in a
basic factorial design (Baayen, 2010). Given two correlated
variables A and B, the residuals from a linear regression of A
onto B will be uncorrelated with B and highly correlated with A.
Using this procedure, we residualized variables when their
Pearson correlation coefficient was higher than .15. All
residualized variables were uncorrelated with the target variable
and correlated highly with the original variable (r>.96).
In addition, model-fitting procedures often have problems
with regression designs containing many variables. Under such
circumstances, the model-fitting procedure may fail to converge.
In our case, the full model that had the appropriate fixed-effect
and random-effect structure failed to converge. Under such
circumstances further reduction of the complexity of the statistical model is recommended (Barr, Levy, Scheepers, & Tily,
2013). Specifically, we distinguished in the model between those
variables related to the compound constituent frequency and
those related to the compound constituent family size. We opted
to investigate the impact of these two sets of variables in analyses
of two distinct statistical models. Furthermore, because there
was a high degree of correlation between the constituent measures related to family size, and because the main focus of the
study was not on distinguishing between these individual measures, a principal component analysis was conducted on all
constituent family size variables. We extracted only the first
component that captured the maximum amount of variance.
We refer to the first principal components corresponding to the
compound’s left and right constituent family size measures as the
left and right principal components. The left and right principal
components captured 70 % and 72 % of variance, respectively.
In sum, variables were centered and residualized in order to
reduce problems of collinearity. In addition, problems with
convergence were addressed by implementing a principal
Mem Cogn
component analysis and by analyzing the data with two different statistical models. The correlation matrix that lists the
relevant predictors for the analyses reported below is presented in Table 2. Note that this matrix reports correlations before
and after the residualization procedures.
Analysis 1: Basic frequency measures
The first analysis focused on the effects of basic frequency
measures. The statistical model included the critical variables
surface frequency, left and right constituent lemma frequency,
all control variables (see above), block (before and after
familiarization), and the interaction of block with all critical
variables. The random-effect structure included by-participant
and by-item random intercepts, by-participant random slopes
for the three frequency variables, and by-item random slopes
for the variable block. The by-participant random-effect structure did not include a parameter that estimated the correlation
between the random variables. More complex random-effect
structures led to models that failed to converge.2
The results from the mixed effect analysis of the full model
is presented in the top-panel of Table 3. The heading “Full
model” lists the beta-coefficients and t-values corresponding
to an analysis in which all predictors are present simultaneously in the model. These results reveal important information
about the sign of the various effects. The heading “Model
comparisons” provides the significance value of the predictor.
The model comparisons were performed using the likelihood
ratio test, where models with and without predictors were
compared while the random-effects structure was kept unchanged (Barr et al., 2013). A graphical presentation of the
results from this analysis is presented in Fig. 1.
The results revealed an effect of block (Fig. 1a), where
latencies became faster in the second block. There was also an
effect of surface frequency in block 1 (Fig. 1b), where latencies became shorter with increasing frequency values. The
effect of surface frequency interacted with block. In addition,
there was an effect of trial (Fig. 1c), where mean naming
latencies became longer along the course of the experiment.3
In addition, there were effects of familiarity (Fig. 1d), name–
image agreement (Fig. 1e), and visual complexity (Fig. 1f).
Crucially, there were no effects of the compound’s left and
right constituent lemma frequencies.4
We further examined the interaction of block and surface
frequency by inspecting the surface frequency effect
separately in block 1 and 2. The statistical models in
these analyses were identical to the model above, except
that the predictor block and all its interactions were
removed. There was an effect of surface frequency in
block 1, t(2699) = −2.65, χ2(1) = 7.39, p < .0066, and in block 2,
albeit smaller in size, t(3741) = −2.49, χ2(1) = 6.51, p < .0108.
Analysis 2: Constituent family size measures
In this analysis, we considered the impact of the variables
related to the compound’s constituent family size. Critical
predictors in this statistical model were the compound’s surface frequency and its left and right principal components. As
before, we also included all control variables, block, and the
interaction with block and the critical variables. The randomeffects structure was similar to that of the first analysis. There
were by-participant and by-item random intercepts, byparticipant random slopes for surface frequency and the two
principal components, and by-item random slopes for the
variable block. More complex random-effect structures failed
to converge (see footnote 2).
The results from the mixed-effect analysis of the full model
is presented in the bottom panel of Table 3. The results
paralleled those found in the first analysis. There was an effect
of block, an effect of surface frequency that interacted with
block, an effect of trial, and effects of familiarity, name–image
agreement, and visual complexity. As in the first analysis,
there were no effects of the left and right principal components
related to the compound’s constituent family size measures.
The interaction of surface frequency with block was further
examined. As in the first analysis, there was an effect of
surface frequency in block 1, t(2699) = −2.83, χ2(1) = 8.38,
p < .0039, and a smaller sized effect in block 2, t(3741) =
−2.50, χ2(1) = 6.57, p < .0104.
Discussion
2
Note that frequency is manipulated between items, and hence, no byitem random slopes for frequency are possible. Furthermore, including
the interaction of the frequency measures with block in the by-participant
random-slopes leads to a model that is unidentified; each participant
rarely contributes more than a single data point for each frequency value
in a given block. These issues contributed to the failure of converge of a
maximal model of random-effects structure (Barr et al., 2013). These
concerns also hold for analysis 2.
3
The reader may be surprised to observe that while latencies became
shorter with increasing block, latencies became longer with increasing
trial. However, this is because trial indexes the position of a trial within a
block, and not across blocks. Hence, although latencies decreased with
increasing block number, within a block, latencies increased with increasing trial number.
Picture-naming latencies in Experiment 1 were sensitive to the compound’s surface frequency, but not to
any of the measures associated with the compound’s
constituents (i.e., lemma frequency, cumulative frequency, family size, positional frequency, positional
entropy, complement frequency, derivational entropy).
In addition, naming latencies were sensitive to the
4
Additional analyses using cumulative constituent frequency instead of
lemma frequency yielded identical results.
Mem Cogn
Table 2 Correlation matrix reporting Pearson’s correlation coefficients between the variables used in the statistical models of Experiment 1
Variables
Fam
NIA
VC
S.freq
Fam
1.00
.15
-.13
.00
NIA
VC
.15
-.13
1.00
.02
.02
1.00
.03
.03
S.freq
.16
.06
.02
1.00
L.lem
R.lem
.16
.15
-.03
.08
-.22
.06
L.PC
R.PC
-.11
.11
.05
.14
.10
-.01
L.lem
R.lem
L.PC
R.PC
-.03
.00
-.11
.11
-.05
.00
.06
.09
.05
.10
.14
-.01
.02
.00
-.05
.00
.19
.16
1.00
-.11
-.12
1.00
-.77
.13
-.14
.83
-.26
.23
-.79
-.11
.12
.84
1.00
.17
.17
1.00
Note. Below the diagonal (gray background), correlations before residualizing and centering; above the diagonal (white background), correlations after
residualizing and centering. Note the absence of any correlation between the surface frequency and relevant constituent variables after residualization. Fam =
familiarity, NIA = name–image agreement, VC = visual complexity, S.freq = surface frequency, L.lem = left constituent lemma frequency, R.lem = right
constituent lemma frequency, L.PC = left (modifier) principal component, R.PC = right (head) principal component. Please note that the correlations between
constituent lemma frequency and principal components were not residualized, given that it was not our concern to distinguish between two types of variables,
and were analyzed in separate models. Note also that although a single surface frequency variable is presented here, two different residualized versions were
used for the analyses of basic frequency and family size measures.
picture’s familiarity, name–image agreement, and visual
complexity measures (e.g., Ellis & Morrison, 1998).
A potential concern with these results is that they rely
exclusively on the use of frequency measures obtained from
CELEX. Recently, researchers have argued that measures
obtained from other sources, such as SUBTLEX, provide
more accurate estimates of word frequency than does CELEX
(e.g., Brysbaert & New, 2009). Consequently, one might argue
that the failure to obtain effects for the constituent frequency
measures in Experiment 1 reflects the specific CELEX measures used in the experiment. To address this issue, we
reanalyzed the data using surface and constituent frequency
measures obtained from SUBTLEX. The results obtained in
this reanalysis were identical to those obtained in Experiment
1, except that we did not find an interaction between block and
surface frequency.
These results establish the reliability of the findings by
Janssen et al., (2008). In picture naming, compound word
production yields a surface frequency effect but no constituent
effects. In addition, the frequency effect was smaller after
familiarization (in block 2), which rules out the possibility
that the familiarization procedure caused the surface frequency effect previously observed by Janssen et al. (2008).
As we discussed above, this pattern of results obtained in
the picture-naming task is at odds with that found in other
language production and comprehension tasks. For example,
constituent effects have been robustly found in lexical decision tasks (e.g., Kuperman et al., 2009). Our main goal here
was to rule out the possibility that the contrasting constituent
effects observed between picture naming and lexical decision
tasks reflect methodological differences. In Experiment 2 we
examined this possibility directly by extracting lexical decision latencies for the compounds used in Experiment 1 from
the English Lexicon Project (Balota et al., 2007).
Experiment 2: Lexical decision
Method
Materials
The English Lexicon Project contains a sample of over 40,000
words for which lexical decision times are available. These
lexical decision times were collected in subsets from a large
number of participants (>800). The participants were students
at universities from various regions of the U.S. For the lexical
decision task, the word stimuli were selected from Francis and
Kučera and CELEX, and the pronounceable nonword stimuli
were created by changing one or two letters in a corresponding
target word (see Balota et al., 2007, for more details).
Lexical decision times were available for a subset of 95 words
from the original set of 150 words (63 %). In this subset, the
average surface frequency was higher than in the original set,
t(243) = 3.49, p < .001, as was the left positional frequency,
t(243) = 3.18, p < .002. No other variables in the subset differed
statistically from those in the original, indicating that the subset is
a representative sample of the compounds used in Experiment 1.
For the lexical properties of the items in this subset, see Table 1.
Analysis and results
Given that our latencies from the English Lexicon Project
were item based (i.e., averaged over participants), we conducted a standard item-based ordinary least squares regression
analysis. As before, to combat collinearity, we centered and
residualized variables with a Pearson correlation coefficient
>.15. The correlation matrix of the variables before and after
residualization is presented in Table 4. We conducted the two
analyses that paralleled those of Experiment 1.
Mem Cogn
Table 3 Analysis of the picture naming latencies of Experiment 1 using mixed effect modeling
Full Model
Model Comparisons
Analysis
Variable
β (SE)
t(6445)
χ2(1)
p
Analysis 1
Intercept
Block = 1
Surface freq
Left lemma freq
Right lemma freq
Trial
Familiarity
Name-image ag
Visual complexity
Plosiveness = yes
Fricative = yes
Voicing = yes
Block:surface freq
Block:left lemma freq
1,020.05 (18.73)
−239.26 (6.86)
−19.92 (6.77)
2.67 (7.23)
−3.52 (7.23)
.29 (0.09)
−25.65 (7.87)
−91.27 (12.73)
16.33 (8.17)
−25.90 (15.52)
15.01 (17.18)
2.76 (13.48)
11.55 (4.08)
−.79 (4.22)
54.46
−34.87
−2.94
.37
−.49
3.23
−3.26
−7.17
2.00
−1.67
.87
.21
2.83
−.19
468.04
1,100.42
8.99
.14
.25
13.43
14.02
47.10
4.22
2.95
.82
.05
7.98
.03
.0000
.0000
.0027
.7039
.6153
.0002
.0002
.0000
.0400
.0861
.3658
.8317
.0047
.8570
1.02
1.04
.3077
1,022.21 (18.93)
−239.33 (6.86)
−21.38 (7.06)
−.63 (5.65)
54.01
−34.90
−3.03
−.11
463.87
1,102.35
9.51
.01
.0000
.0000
.0020
.9086
−2.08 (5.27)
.28 (0.09)
−23.18 (8.08)
−91.91 (12.75)
16.84 (8.16)
−26.99 (15.84)
12.74 (17.34)
1.59 (13.59)
14.91 (4.21)
−2.51 (3.21)
3.34 (3.19)
−.39
3.20
−2.87
−7.21
2.06
−1.70
.73
.12
3.54
−.78
1.04
.16
10.27
8.57
47.49
6.07
3.07
2.17
.01
12.50
.62
1.09
.6863
.0014
.0034
.0000
.0138
.0796
.1403
.9034
.0004
.4304
.2969
Block:right lemma freq
Analysis 2
Intercept
Block = 1
Surface freq
Left PC
Right PC
Trial
Familiarity
Name–image ag
Visual complexity
Plosiveness = yes
Fricative = yes
Voicing = yes
Block:surface freq
Block:left PC
Block:right PC
4.38 (4.29)
Note. Analysis 1 refers to the basic frequency measures, and analysis 2 refers to the family size measures. Full model refers to the results from the analysis
when all predictors are simultaneously present in the model. Model comparison refers to the results from likelihood ratio tests, where a model with and
without the predictor are compared. Degrees of freedom were determined by subtracting the number of fixed effect predictors from the total number of
data points (Baayen, 2008). Note the effect of surface frequency and the absence of effects of constituent measures. Name–image ag = name–image
agreement; freq = frequency; PC = principal component.
For the first analysis we considered the variables surface frequency and the left and right constituent lemma
frequency. As can be seen in Table 5 and Fig. 2, lexical
decision latencies were sensitive to the compound’s surface frequency (Fig. 1a) and its left and right constituent
lemma frequency (Fig. 1b, c).
In the second analysis, we included the variables
surface frequency and the left and right constituent
principal components associated with a compound’s
constituent family size. This analysis revealed effects
of the surface frequency (Fig. 1d) and of the left and
the right constituent principal components (Fig. 1e, f).
Discussion
The results of Experiment 2 contrast sharply with those of
Experiment 1. Whereas in Experiment 1, picture-naming
latencies were affected by the compound’s surface frequency effect, but not its constituent measures, in Experiment 2, the lexical decision latencies were sensitive to
0.2 0.4
−3
−1
1 2 3
2
0.006
0.003
density
0
−1
0
1
2
0.10 0.20
density
800
400
0.00
0.2
0.4
−2
0.30
F
0.0
800
400
3
Familiarity
50
Trial
1200
0.10
0.20
1
0.00
800
400
0
0.000
−50
E
1200
D
−2
1200
800
0.20
4
Surface Frequency
1200
Block
400
400
−0.2
0.10
1200
block1
block2
0.00
1200
−0.6
Naming latencies (in ms)
C
800
B
800
A
400
Naming latencies (in ms)
Mem Cogn
−3
Name image agreement
−1
0
1
2
3
Visual Complexity
Fig. 1 Graphical presentation of the picture-naming results of Experiment 1. Each panel (a–f) shows the effect of a significant predictor on
naming latencies as a function of a predictor’s values on the x-axes. The
effect of each predictor is adjusted for the effects of the other predictors.
Naming latencies are shown on the left y-axes (in milliseconds) in black.
Also plotted are the relative occurrences of each value of a continuous
predictor (by means of a smoothed histogram function) on the right y-axes
in gray
the compound’s surface frequency and to its left and right
constituent measures.
Two further methodological concerns need to be addressed
before we can proceed to a general interpretation of the results.
First, Experiment 2 contained a subset of the items used in
Experiment 1. Therefore, one could argue that the contrasting
results between the two experiments were caused by these
different sets of items. However, a reanalysis of Experiment 1
with the items of Experiment 2 did not substantially change
the pattern of results of Experiment 1. Specifically, as before,
surface frequency effects were obtained in the absence of
effects for the compound’s constituent measures (see Table 6
for details).
Second, different types of analyses were used between Experiments 1 and 2. Experiment 1 relied on a
mixed-effects technique with crossed random effects for
participants and items, while Experiment 2 relied on an
item-based ordinary least squares regression. To rule out
that this difference in the type of analyses impacted the
results, we reanalyzed the results of Experiment 1 with
an ordinary least squares method. This was achieved by
averaging the naming latencies across participants,
thereby obtaining item-based RT averages. This least
squares analysis produced the same results as those
found with the mixed-effects method, except that no
interaction between block and surface frequency was
found (see Table 7 for details).
Table 4 Correlation matrix reporting Pearson’s correlation coefficients
between the variables used in the statistical models of Experiment 2
reported below.
Variables
S.freq
L.lem
R.lem
L.PC
S.freq
1.00
-.07
.00
-.12
R.PC
.00
L.lem
.08
1.00
.03
-.81
-.05
R.lem
.33
.03
1.00
.02
.82
L.PC
-.07
-.81
.02
1.00
.15
R.PC
.34
-.05
.82
.15
1.00
Note. Below the diagonal (gray background), correlations before
residualizing and centering; above the diagonal (white background),
correlations after residualizing and centering. Note the absence of any
correlation between the surface frequency and relevant constituent variables after residualization. S.freq = surface frequency, L.lem = left constituent lemma frequency, R.lem = right constituent lemma frequency,
L.PC = left (modifier) principal component, R.PC = right (head) principal
component; As before, correlations between constituent lemma frequency
and principal components were not residualized, and two different
residualized versions of surface frequency were used for the analyses of
basic frequency and family size measures.
General discussion
In Experiment 1, picture-naming latencies were sensitive
to the compound’s surface frequency, picture familiarity,
and name–image agreement properties, but not to its
constituent measures. In Experiment 2, lexical decision
Mem Cogn
Table 5 Analysis of the lexical decision latencies of Experiment 2 using
ordinary least squares
Analysis
Variable
β (SE)
t(91)
p
Analysis 1
Intercept
Surface frequency
Left lemma frequency
Right lemma frequency
718.37 (7.95)
−12.46 (5.30)
−15.31 (5.76)
−21.66 (4.98)
90.30
−2.35
−2.66
−4.35
.0001
.0208
.0094
.0001
Analysis 2
Intercept
Surface frequency
Left PC
Right PC
718.37 (7.87)
−12.85 (5.28)
18.20 (4.59)
−14.93 (3.80)
91.23
−2.43
3.96
−3.93
.0001
.0171
.0002
.0002
Note. Analysis 1 refers to the analysis of the frequency variables, and
analysis 2 refers to the analysis of the family size variables. Note the
presence of both surface and constituent effects. PC = principal component.
latencies were sensitive to the compound’s surface frequency, left and right constituent frequency, and family
size measures.
These results confirm the contrasting effects of constituent
measures in the picture-naming and lexical decision tasks
previously found in the literature. For example, Janssen
et al. (2008) reported surface frequency but not constituent
effects in a picture-naming task with compound words, whereas Juhasz et al. (2003; Kuperman et al., 2009) reported surface
frequency and constituent effects in a lexical decision task.
These contrasting constituent effects were also obtained in our
picture-naming and lexical decision tasks, where the same
methodologies (materials, language, analyses) were used.
Thus, methodological differences alone cannot account for
the observed contrast.
Instead, we propose that this pattern of contrasting constituent effects arises due to differences in the degree to which the
input representation in a task transparently contains the compound’s constituents. Consider first word comprehension
tasks such as lexical decision. In this task, the input representation is an orthographic or phonological string. This input
string transparently contains the compound word’s constituents, meaning that they are directly recoverable from the
physical signal. For such input representations, it is assumed
that they lead to the activation of the compound’s constituents
in the lexicon, which in turn leads to the observed constituent
effects. By contrast, in the picture-naming task, a semantic
input representation drives word retrieval (e.g., Potter &
Faulconer, 1975). Unlike the orthographic or phonological
input representations in word comprehension tasks, a semantic input representation does not transparently contain a compound’s constituents. For such input representations, it is
assumed that they do not directly activate the constituents in
the lexicon, leading to the absence of constituent effects.5
Thus, the presence of constituent effects in a task depends
on the degree to which the input representation in a task
transparently contains the compound’s constituents.
The absence of constituent effects in the picture-naming
task reported here contrasts with the presence of such effects
in other word production tasks, such as the response association and picture–word interference tasks (Bien et al., 2005;
Roelfs, 1996; Zwitserlood et al., 2000). Although our study
was not designed to resolve this issue, we discuss two possible
reasons for this inconsistency. First, the observed discrepancy
could be the result of the different languages used: Whereas
constituent effects have been reported in studies using German
and Dutch speakers, no constituent effects have been reported
in studies using English and Chinese speakers. Crosslinguistic differences may influence compound representation
and account for the contrasting results in the reported studies.
However, note that this explanation assumes that compound
representation is more similar between English and Chinese
than between English and Dutch, which seems unlikely (see
Janssen et al., 2008, for a similar argument). A more likely
possibility is that compound processing differs between tasks,
where some tasks emphasize the processing of the compound’s constituents, whereas other tasks do not. Consider,
for example, the observation that compound word production
yields robust surface frequency effects in the picture-naming
task, but not in the response association task (Bien et al., 2005,
exp4). This contrast in the presence of the surface frequency
effect clearly points to underlying differences between the two
tasks that may affect the processing of compound words.
Future studies should be geared towards resolving this issue.6
Other studies have also used the picture-naming task to
investigate the production of complex words other than compounds. These studies have reported both surface and constituent effects in the production of singular and plural inflected
nouns (Baayen, Levelt, Schreuder, & Ernestus, 2007) and
regular and irregular verbs (Tabak, Schreuder, & Baayen,
2010). These data are fully compatible with the proposal
advanced here, if one further nuance is taken into account.
This concerns the assumption that the nature of the input
representation in a task is determined not only by the task,
but by a combination of the task and the specific type of the
5
Compound constituents may be activated weakly because of shared
semantic representations, just as would be the case for any two semantically related lexical entries.
6
Constituent effects have also been observed in various versions of the
picture–word interference task (Zwitserlood et al., 2000). However, as
was discussed by Janssen et al. (2008), these effects may reflect semantic
and not morphological priming. Specifically, the interpretation of longlag morphological priming effects rests on the absence of long-lag semantic priming effects (Feldman, 2000). However, long-lag semantic
priming effects are reported in the literature (e.g., Becker, Moscovitch,
Behrmann, & Joordens, 1997), thereby undermining a nonsemantic interpretation of long-lag morphological priming effects. See Janssen et al.
(2008) for further discussion of this issue.
Mem Cogn
4
0
2
4
Left constituent PC
0.20
density
0.10
0.10
0.00
density
0.20
900
800
0.00
−2
700
0.10
800
700
0.10
0.00
2
0.20
900
F
600
0.20
900
800
700
600
0
0.00
−4 −2
0
2
4
Right constituent frequency
E
Surface frequency
900
700
600
−6 −4 −2 0
2
Left constituent frequency
D
−2
800
0.30
0.10
0.00
0.20
900
700
4
600
−1
1 2 3
Surface frequency
600
0.10
0.00
700
−3
Decision latencies (in ms)
C
800
800
0.20
900
B
600
Decision latencies (in ms)
A
−6
−4
−2
0
2
4
Right constituent PC
Fig. 2 Graphical presentation of the lexical decision results of Experiment 2, which considered the effects of surface and left and right constituent frequency (top row) and the effects of surface frequency and left and
right constituent PCs (bottom row) in two separate analyses. The effect of
each predictor is adjusted for the effects of the other predictors. Decision
latencies are shown on the left y-axes (in milliseconds) in black, and the
relative occurrences of each value of a continuous predictor are shown on
the right y-axes in gray
complex word that is used in the experiment. In other words, it
is not our proposal that the task per se determines whether the
constituents are activated in the lexicon. Rather, our proposal is
that both the type of complex word and the task together
determine the nature of the input representation and that this,
in turn, determines the activation of the constituents. With
respect to the constituent effects found in the picture-naming
studies with inflected nouns and verbs, we make the plausible
assumption that the semantic representation that underlies plural inflected nouns and verbs transparently contains the complex word’s constituents, leading to the observed constituent
effects. This would then lead to the testable prediction that in a
picture-naming task with compound word production, constituent effects should be modulated by the semantic transparency
of the compound (Libben et al., 2003).
Our results challenge the notion that morphology plays a
central and obligatory role in the retrieval of words from the
lexicon. Models that incorporate this idea have been proposed
in both the word comprehension (e.g., Taft, 2003; Taft &
Ardasinski, 2006), and word production (e.g., Levelt, Roelofs,
& Meyer, 1999) literature. These models assume that a word’s
lexical representation involves a decomposed morphemelevel representation and that access to these morpheme-level
representations is a necessary part of word retrieval. The
contrasting constituent effects in compound word processing
between picture-naming and lexical decision tasks observed
here present problems for models that rely on these assumptions. Specifically, if access to a compound’s constituents
were obligatory for the retrieval of the compound word from
the lexicon, one would expect effects of the compound’s
constituents in all tasks. The observation that constituent
effects were observed in lexical decision, but not in picture
naming, therefore undermines the notion that morphology
plays a central and obligatory role in lexical access.
In the word comprehension literature, alternative models
that reject the idea of obligatory access have been proposed.
Specifically, such models assume that the recognition of morphologically complex words relies on two parallel routes: a
whole-word route in which the access to meaning takes place
on the basis of the whole-word form and a decomposed route
in which the meaning of the complex word is accessed on the
basis of its constituents (e.g., Caramazza et al., 1988;
Schreuder & Baayen, 1997). A dual-route-type architecture
could account for the results observed here by assuming that
the degree to which the decomposed route is activated depends on the degree to which the input representation of a task
transparently contains the compound’s constituents. In the
lexical decision task, the decomposed route is strongly activated on the basis of the transparent input representation,
leading to the observed constituent effects. In a picturenaming task with compound word stimuli, the decomposed
route is weakly or not activated, leading to the absence of
constituent effects.7
One additional aspect of our data deserves further discussion. In the picture-naming task of Experiment 1, we did not
observe any effects of the measures associated with the compound’s left and right constituent family size. This is surprising, considering that family size effects have been interpreted
Mem Cogn
Table 6 Model statistics of the variables in a reanalysis of Experiment 1 with the 95 items used in Experiment 2
Full Model
Model Comparisons
Analysis
Variable
β (SE)
t(4236)
χ2(1)
p
Analysis 1
Intercept
Block = 1
Surface freq
Left lemma freq
Right lemma freq
Trial
Familiarity
Name–image ag
Visual complexity
Plosiveness = yes
Fricative = yes
Voicing = yes
Block:surface freq
Block:left lemma freq
966.02 (25.32)
−232.20 (8.34)
−15.58 (9.69)
−2.82 (11.18)
−3.02 (9.11)
.37 (0.11)
−26.76 (10.11)
−92.14 (15.66)
13.22 (10.44)
1.08 (20.03)
48.64 (23.55)
12.32 (18.02)
9.97 (5.42)
−.49 (6.17)
38.14
−27.81
−1.61
−.25
−.33
3.51
−2.64
−5.88
1.27
.05
2.06
.68
1.84
−.08
275.67
701.38
4.01
.07
.12
12.11
7.52
33.72
1.77
.00
5.83
.67
3.41
.01
.0000
.0000
.0452
.7912
.7279
.0005
.0061
.0000
.1829
.9544
.0157
.4135
.0648
.9301
1.29
2.76
.0969
964.07 (25.15)
−232.00 (8.34)
−18.18 (9.68)
−5.21 (8.97)
38.33
−27.80
−1.88
−.58
276.78
700.87
3.30
.81
.0000
.0000
.0692
.3668
2.63 (6.86)
.37 (0.11)
−28.48 (10.20)
−92.58 (15.53)
12.96 (10.31)
−.37 (19.63)
52.45 (23.72)
12.82 (17.78)
12.59 (5.45)
.93 (4.87)
−1.46 (4.01)
.38
3.45
−2.79
−5.96
1.26
−.02
2.21
.72
2.31
.19
−.36
.17
11.13
7.77
32.69
1.74
.00
4.74
1.03
5.73
.00
.00
.6787
.0008
.0053
.0000
.1874
.9999
.0294
.3108
.0167
.9999
.9999
Block:right lemma freq
Analysis 2
Intercept
Block = 1
Surface freq
Left PC
Right PC
Trial
Familiarity
Name–image ag
Visual complexity
Plosiveness = yes
Fricative = yes
Voicing = yes
Block:surface freq
Block:left PC
Block:right PC
6.69 (5.18)
Note. Note the absence of any constituent effects. Name–image ag = name–image agreement; freq = frequency; PC = principal component.
not at a morphological, but at a semantic level (e.g., Dijkstra,
Moscoso del Prado Martin, Schulpen, Schreuder, & Baayen,
2005; Moscoso del Prado Martin, Bertram, Häikiö, Schreuder,
& Baayen, 2004). Given that the picture-naming task is
sensitive to the processing of semantic information, one
would have expected family size effects in this task. One
possible explanation of this state of affairs is within the
dual-route view proposed above. One might assume that
the family size effect does not reflect any intrinsic aspect
7
Connectionist models (Baayen, Milin, Durdević, Hendrix, & Marelli,
2011; Rueckl, & Raveh, 1999; Seidenberg & Gonnerman, 2000) also
reject the assumption of obligatory access and, therefore, could, in principle, also account for the results presented here.
of the meaning of a complex word but that its effects
arise from the decomposed processing of the complex
word’s constituents. Under such an assumption, one
would expect no effect of a compound’s constituent
family size in a picture-naming task, but one would
expect such effects in other tasks in which the input
representation allows for decomposed retrieval. Such a
proposal seems borne out by the data, where family size
effects have been found in response association (Bien
et al., 2005) and comprehension tasks such as lexical
decision and naming (Juhasz & Berkowitz, 2011). Further targeted studies should confirm this interpretation.
Finally, the present results add to a growing body of data
that show a failure to detect constituent effects in compound
Mem Cogn
Table 7 Model statistics of the variables in a reanalysis of Experiment 1
using the ordinary least squares regression method used in Experiment 2
Full Model
Analysis
Variable
Analysis 1 Intercept
Block = 1
Surface freq
Left lemma freq
Right lemma freq
Familiarity
Name–image ag
Visual complexity
Plosiveness = yes
Fricative = yes
Voicing = yes
Block:surface freq
Block:left lemma freq
Block:right lemma freq
Analysis 2 Intercept
Block = 1
Surface freq
Left PC
Right PC
Familiarity
Name–image ag
Visual complexity
Plosiveness = yes
Fricative = yes
β (SE)
1,059.95 (16.82)
t(285)
p
63.00 .0001
−264.81 (19.82) −13.36 .0001
−20.86 (6.02)
−3.47 .0007
4.18 (6.26)
0.67 .5052
−5.99 (6.40)
−0.94 .3494
−29.72 (7.02)
−4.23 .0001
−99.45 (11.46) −8.67 .0001
17.92 (7.28)
2.46 .0145
−22.65 (13.88) −1.63 .1039
19.78 (15.24)
1.30 .1953
−0.49 (12.00) −0.04 .9669
11.03 (12.03)
.92 .3602
−3.34 (12.35) −0.27 .7868
7.53 (12.69)
.59 .5534
1,062.17 (17.01) 62.44 .0000
−264.82 (19.79) −13.38 .0001
−22.93 (6.29)
−3.65 .0003
−2.38 (4.83)
−0.49 .6224
−2.56 (4.67)
−0.55 .5842
−27.19 (7.19)
−3.78 .0002
−100.15 (11.47) −8.73 .0001
18.23 (7.27)
2.51 .0127
−24.14 (14.13) −1.71 .0888
17.23 (15.38)
1.12 .2635
Voicing = yes
−2.11 (12.09)
Block:surface freq
Block:left PC
Block:right PC
16.68 (12.47)
.86 (9.30)
4.88 (9.13)
−0.18 .8615
1.33 .1823
.09 .9261
.54 .5934
Note. Note that this analysis does not require model comparisons to
obtain p-values and that the variable trial could not be included. Note
also the persistent surface frequency effect in the absence of a constituent
effect. Name–image ag = name–image agreement; freq = frequency; PC =
principal component.
word production. Besides the data of Janssen et al. (2008),
two recent studies have also failed to find effects of constituents in compound word production. First, Cohen-Goldberg
(2013) examined the duration of compound word latencies in
natural speech selected from the Buckeye corpus (Pitt et al.,
2007). Using regression analyses, Cohen-Goldberg found that
compound word duration latencies were predicted by the
compound’s surface frequency, but not by its constituent
frequency. Likewise, Jacobs and Dell (in press) failed to
obtain a second constituent priming effect during compound
word production. Specifically, phonological priming effects
patterned similarly for compounds and bisyllabic
monomorphemic words and differed from those found for
two-word utterances. Both Cohen-Goldberg and Jacobs and
Dell have interpreted these data in terms of a model that
assumes whole-word representations for compounds at the
phonological level. These independently collected data points
instill further confidence that the failure to detect the constituent effects in the presenst study is a real null effect.
To conclude, the primary motivation of our study was to
resolve a surprising discrepancy in the literature: The observation that whereas the majority of tasks reveal constituent
effects in compound processing, the picture-naming task does
not. Our results establish that these contrasting constituent
effects cannot be explained by methodological differences
alone. Instead, we propose that these contrasts are more likely
the result of task-specific differences in the degree to which
the input representation in a given task transparently contains
the compound’s constituents. Our results are not readily accommodated by models that assume that access to a compound’s constituents is an obligatory aspect of compound
word retrieval (Levelt et al., 1999; Taft, 2003) but find a more
straightforward explanation in models that assume parallel
dual routes to constituent access (Caramazza et al., 1988;
Schreuder & Baayen, 1997), where the activation of the
decomposed route depends on the nature of the input representation in the task. An avenue for future research would be
to examine the role of semantic transparency in the production
of compounds in the picture-naming task.
Acknowledgments This research was supported by Grant RYC-201108433 from the Spanish Ministry of Economy and Competitiveness to
N.J. and by Grant R01DC006842 from the National Institute on Deafness
and Other Communication Disorders and the Societã Mente/Cervello,
Fondazione Cassa di Risparmio di Trento e Rovereto to A.C.
References
Andrews, S. (1986). Morphological influences on lexical access: Lexical or
nonlexical effects? Journal of Memory and Language, 25(6), 726–740.
Baayen, R. (2008). Analyzing linguistic data: A practical introduction to
statistics using R. Cambridge: Cambridge Univ Pr.
Baayen, R. (2010). A real experiment is a factorial experiment. The
Mental Lexicon, 5(1), 149–157.
Baayen, R., Piepenbrock, R., & van Rijn, H. (1993). The {CELEX}
lexical data base on {CD – ROM}.
Baayen, R., Levelt, W., Schreuder, R., & Ernestus, M. (2007). Paradigmatic
structure in speech production. In Proceedings from the Annual Meeting
of the Chicago Linguistic Society (pp. 1–29, vol 43). Chic Ling Society.
Baayen, R., Milin, P., Đurđević, D., Hendrix, P., & Marelli, M. (2011). An
amorphous model for morphological processing in visual comprehension based on naive discriminative learning. Psychological Review,
118(3), 438–481.
Balota, D., Cortese, M., Sergent-Marshall, S., Spieler, D., & Yap, M.
(2004). Visual word recognition of single-syllable words. Journal of
Experimental Psychology: General, 133(2), 283–316.
Mem Cogn
Balota, D., Yap, M., Hutchison, K., Cortese, M., Kessler, B., Loftis, B.,
… Treiman, R. (2007). The English Lexicon Project. Behavior
Research Methods, 39(3), 445–459.
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects
structure for confirmatory hypothesis testing: Keep it maximal.
Journal of Memory and Language, 68(3), 255–278.
Becker, S., Moscovitch, M., Behrmann, M., & Joordens, S. (1997). Longterm semantic priming: A computational account and empirical
evidence. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 23(5), 1059–1082.
Bertram, R., & Hyönä, J. (2003). The length of a complex word modifies
the role of morphological structure: Evidence from eye movements
when reading short and long finnish compounds. Journal of
Memory and Language, 48(3), 615–634.
Bien, H., Levelt, W., & Baayen, R. (2005). Frequency effects in compound production. Proceedings of the National Academy of Sciences
of the United States of America, 102(49), 17876–17881.
Brysbaert, M., & New, B. (2009). Moving beyond Kučera and Francis: A
critical evaluation of current word frequency norms and the introduction of a new and improved word frequency measure for american english. Behavior Research Methods, 41(4), 977–990.
Caramazza, A., Laudanna, A., & Romani, C. (1988). Lexical access and
inflectional morphology. Cognition, 28(3), 297–332.
Cohen-Goldberg, A. M. (2013). The measure of a (post)man:
Investigating the lexical representation of compounds in production.
Talk presented at the XIth International Symposium of
Psycholinguistics. Spain: Tenerife.
De Jong, N., Feldman, L., Schreuder, R., Pastizzo, M., & Baayen, R.
(2002). The processing and representation of Dutch and English
compounds: Peripheral morphological and central orthographic effects. Brain and Language, 81(1–3), 555–567.
Diependaele, K., Grainger, J., & Sandra, D. (2012). Derivational morphology and skilled reading: An empirical overview.
Dijkstra, T., Moscoso del Prado Martín, F., Schulpen, B., Schreuder, R.,
& Baayen, R. (2005). A roommate in cream: Morphological family
size effects on interlingual homograph recognition. Language and
Cognitive Processes, 20(1–2), 7–41.
Dohmes, P., Zwitserlood, P., & Bölte, J. (2004). The impact of semantic
transparency of morphologically complex words on picture naming.
Brain and Language, 90(1–3), 203–212.
Duñabeitia, J., Perea, M., & Carreiras, M. (2007). Do transposed-letter
similarity effects occur at a morpheme level? evidence for morphoorthographic decomposition. Cognition, 105(3), 691–703.
Ellis, A., & Morrison, C. (1998). Real age-of-acquisition effects in lexical
retrieval. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 24(2), 515–523.
Feldman, L. B. (2000). Are morphological effects distinguishable from
the effects of shared meaning and shared form? Journal of
Experimental Psychology: Learning, Memory, and Cognition,
26(6), 1431–1444.
Forster, K., & Forster, J. (2003). Dmdx: A windows display program with
millisecond accuracy. Behavior Research Methods, 35(1), 116–124.
Glaser, W. (1992). Picture naming. Cognition, 42(1–3), 61–105.
Gumnior, H., Bölte, J., & Zwitserlood, P. (2006). A chatterbox is a box:
Morphology in german word production. Language and Cognitive
Processes, 21(7–8), 920–944.
Jacobs, C. L., & Dell, G. S. (in press). “hotdog”, not “hot dog”: The
phonological planning of compound words. Language, Cognition,
and Neuroscience
Janssen, N., Bi, Y., & Caramazza, A. (2008). A tale of two frequencies:
Determining the speed of lexical access for mandarin chinese and
english compounds. Language and Cognitive Processes, 23(7–8),
1191–1223.
Janssen, N., Pajtas, P., & Caramazza, A. (2011). A set of 150
pictures with morphologically complex english compound
names: Norms for name agreement, familiarity, image
agreement, and visual complexity. Behavior Research
Methods, 43, 478–490.
Juhasz, B. (2008). The processing of compound words in English: Effects
of word length on eye movements during reading. Language and
Cognitive Processes, 23(7–8), 1057–1088.
Juhasz, B., & Berkowitz, R. (2011). Effects of morphological families on
english compound word recognition: A multitask investigation.
Language and Cognitive Processes, 26(4–6), 653–682.
Juhasz, B., Starr, M., Inhoff, A., & Placke, L. (2003). The effects of
morphology on the processing of compound words: Evidence from
naming, lexical decisions and eye fixations. British Journal of
Psychology, 94(2), 223–244.
Koester, D., & Schiller, N. O. (2008). Morphological priming in overt
language production: Electrophysiological evidence from Dutch.
NeuroImage, 42(4), 1622–1630.
Kuperman, V., Schreuder, R., Bertram, R., & Baayen, R. (2009). Reading
polymorphemic dutch compounds: Toward a multiple route model of
lexical processing. Journal of Experimental Psychology: Human
Perception and Performance, 35(3), 876–895.
Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999). A theory of lexical
access in speech production. Behavioral and Brain Sciences, 22, 1–38.
Libben, G., Gibson, M., Yoon, Y., & Sandra, D. (2003). Compound
fracture: The role of semantic transparency and morphological
headedness. Brain and Language, 84(1), 50–64.
Moscoso del Prado Martín, F., Bertram, R., Häikiö, T., Schreuder, R., &
Baayen, R. (2004). Morphological family size in a morphologically rich
language: The case of finnish compared with Dutch and Hebrew.
Journal of Experimental Psychology: Learning, Memory, and
Cognition, 30(6), 1271–1278.
Pitt, M., Dilley, L., Johnson, K., Kiesling, S., Raymond, W., Hume, E., &
Fosler-Lussier, E. (2007). Buckeye corpus of conversational speech
(2nd release). Columbus, OH: Department of Psychology, Ohio
State University.
Pollatsek, A., Hyönä, J., & Bertram, R. (2000). The role of morphological
constituents in reading finnish compound words. Journal of
Experimental Psychology: Human Perception and Performance,
26(2), 820–833.
Potter, M., & Faulconer, B. (1975). Time to understand pictures and
words. Nature, 253, 437–438.
Roelofs, A. (1996). Morpheme frequency in speech production: Testing
weaver. Yearbook of Morphology, 1996, 135–154.
Rueckl, J., & Raveh, M. (1999). The influence of morphological regularities on the dynamics of a connectionist network. Brain and
Language, 68(1–2), 110–117.
Schreuder, R., & Baayen, R. (1997). How complex simplex words can be.
Journal of Memory and Language, 37, 118–139.
Seidenberg, M., & Gonnerman, L. (2000). Explaining derivational morphology as the convergence of codes. Trends in Cognitive Sciences,
4(9), 353–361.
Shannon, C. (2001). A mathematical theory of communication. ACM
SIGMOBILE Mobile Computing and Communications Review, 5(1),
3–55.
Snodgrass, J., & Vanderwart, M. (1980). A standardized set of 260
pictures: Norms for name agreement, image agreement, familiarity,
and visual complexity. Journal of Experimental Psychology:
Human Learning and Memory, 6(2), 174–215.
Tabak, W., Schreuder, R., & Baayen, R. (2010). Producing inflected
verbs: A picture naming study. The Mental Lexicon, 5(1), 22–
46.
Taft, M. (2003). Morphological representation as a correlation between
form and meaning. Reading Complex Words, 1, 113–137.
Taft, M., & Ardasinski, S. (2006). Obligatory decomposition in reading
prefixed words. The Mental Lexicon, 1(2), 183–199.
Taft, M., & Forster, K. (1976). Lexical storage and retrieval of
polymorphemic and polysyllabic words. Journal of Verbal
Learning and Verbal Behavior, 15(6), 607–620.
Mem Cogn
Zwitserlood, P., Bölte, J., & Dohmes, P. (2000). Morphological effects on
speech production: Evidence from picture naming. Language and
Cognitive Processes, 15(4–5), 563–591.
Zwitserlood, P., Bölte, J., & Dohmes, P. (2002). Where and how morphologically complex words interplay with naming pictures. Brain
and Language, 81(1–3), 358–367.